SlideRuler
A single-slide scanner bias calibration method for pathology foundation models. It learns a transfer map to align feature spaces across scanners without modifying frozen foundation model parameters, performing cross-scanner feature correction using only a single calibration slide.
Timeline (last 90 days)
The arXiv paper "One-Slide Calibration of Pathology Foundation Models" proposes the SlideRuler method for calibrating scanner-induced bias in pathology foundation models
SlideRuler does not modify the frozen foundation model, but instead uses intra-slide regions as internal controls to estimate and correct feature drift introduced during acquisition
SlideRuler first learns a transfer map using paired repeat scans, and at inference time calibrates only a single scan, keeping the foundation model parameters unchanged throughout the process
The paper conducted validation on two encoders and five SCORPION scanners
The learned transfer map reduces the average feature distance from target to source by 16.3% to 38.5% compared to the original embeddings
The source-anchored variant proposed in the paper further reduces source feature displacement by 47.7% to 83.6% compared to purely learned transfer, while preserving most of the alignment gains
Across all four evaluation settings (including scanner holdout), same-slide contribution yielded positive effects
SlideRuler's experiments have not yet covered more extreme cross-brand or cross-imaging-modality scenarios (such as brightfield versus fluorescence)
SlideRuler provides improvements at the level of embedding distances, and to what extent these improvements translate into gains in clinical diagnostic accuracy remains to be verified by future research
All Facts (9)
The arXiv paper "One-Slide Calibration of Pathology Foundation Models" proposes the SlideRuler method for calibrating scanner-induced bias in pathology foundation models
50%UnverifiedSlideRuler does not modify the frozen foundation model, but instead uses intra-slide regions as internal controls to estimate and correct feature drift introduced during acquisition
50%UnverifiedSlideRuler first learns a transfer map using paired repeat scans, and at inference time calibrates only a single scan, keeping the foundation model parameters unchanged throughout the process
50%UnverifiedThe paper conducted validation on two encoders and five SCORPION scanners
50%UnverifiedThe learned transfer map reduces the average feature distance from target to source by 16.3% to 38.5% compared to the original embeddings
50%UnverifiedThe source-anchored variant proposed in the paper further reduces source feature displacement by 47.7% to 83.6% compared to purely learned transfer, while preserving most of the alignment gains
50%UnverifiedAcross all four evaluation settings (including scanner holdout), same-slide contribution yielded positive effects
50%UnverifiedSlideRuler's experiments have not yet covered more extreme cross-brand or cross-imaging-modality scenarios (such as brightfield versus fluorescence)
50%UnverifiedSlideRuler provides improvements at the level of embedding distances, and to what extent these improvements translate into gains in clinical diagnostic accuracy remains to be verified by future research
50%